Yishui, reporting from Aofei Temple
QbitAI | WeChat public account QbitAI
Opus 5.5 directly generates a music video from code — and it’s a hit.
Elon Musk practically collapsed on the spot! He not only reposted it multiple times, but also fired off comments like “Whoa!”, “Epic!”, and “This time I truly felt AGI.”
Everyone, experience it directly:
△ Video from Musk’s repost
Characters, lyrics, collage, grain, and transitions pound along to the beat. Over five minutes, your eyes simply can’t keep up.
Still not satisfied?
Then speed-run “Google’s Road to Here” in two minutes. The animation cuts faster and faster, the soundtrack builds higher and hotter, and it makes your blood boil.
△ Video from x@Deedy
Equally rousing is this MV commemorating Steve Jobs.
In just two minutes, from garage startup to changing personal computing, Jobs’s dramatic life is edited by Opus 5.5 into a Silicon Valley superhero film.
23 kinds of transitions race along to a 120 BPM soundtrack — by the end, you almost want to shout along:
One more thing!
△ Video from x@onur ozcan
Then, one more.
A blogger dug out an old acoustic guitar song he recorded in 2011, first used Suno to rearrange it into a Techno electronic version, then had Opus 5.5 pair it with a cyber MV.
Man, it took me straight to the dance floor.
△ Video from x@Jade Q Wang
After watching, netizens are already sweating for professional motion designers:
“With so many beautiful details, I’m starting to feel sorry for professional motion designers.”
Some simply surrender:
“I love this to death.”
Fine — before the video even finishes, humanity has already raised the white flag (doge).
“Donald’s Big Stage: If You’ve Got the Guts, Come On Up”
In this MV battle royale, one challenge is especially hot right now.
The rules are simple:
Take the same song and the same reference video, hand them to Opus 5.5, and have it shoot the whole thing again.
The person who started the chain is a blogger named Donald.
He first saw “Claude Pop,” made by NotinReality, on X, and casually threw the original link and audio to Opus 5.5, asking it to keep the song but rebuild all the visuals from scratch.
The original looks like this:
Donald talked to his computer for 5 minutes, then let go completely.
Opus 5.5 wrote code, made storyboards, matched visuals, added transitions, and worked nonstop for 12 hours.
By the time Donald woke up, a new 2-minute-21-second MV was already sitting in his computer.
After Donald had his fun, he also kindly published the full prompt of about 9,500 characters.
△ Partial screenshot of Donald’s prompt
Netizens saw it: the original song is here, the reference video is here, the construction manual is here — so what are we waiting for?
So masters from all sides began to compete on the same stage.
Blogger Pleometric turned it into a 2D sci-fi apocalypse theme song:
A red-haired protagonist, pixel stage, retro game UI — sometimes it looks like an anime OP, sometimes like accidentally wandering into an arcade.
mexicat switched to a brighter, more playful visual direction, and did not chase a “one-take success”; he took the finished cut and kept having Claude make fine adjustments.
The result also satisfied him greatly, and he exclaimed “pretty cool.”
The most extreme was anabology.
He added Midjourney and moodboard to Donald’s approach, had Opus 5.5 keep working for another 12 hours, and finally hammered out an extra-long version over five minutes.
That is the one that made Musk collapse at the start.
All right, all right — AI always starts large-scale chain games in unexpected places.
First there was the mass “Ghibli-fication” of ChatGPT images, then Google Nano Banana’s photo-editing contest, and now it’s the coding models’ turn to host an MV same-topic contest.
Knock on the table — if you want to enter now, after summarizing Donald’s prompt, here’s roughly how to do it:
1. Give Claude the original song, the reference MV, and the art style you like, and explain what story you want to tell.
2. In Claude Code, choose Opus 5.5 and paste Donald’s prompt. Install Node.js and FFmpeg on your machine; connect Midjourney and Suno when you need images and music.
3. Let it break down lyrics, write storyboards, make animations, hit the beats, and finally export the complete video.
A friendly reminder.
At first I thought this was pretty simple — just throw the reference files to Opus 5.5, right?
But when I flipped through the comments, the tone gradually went off...
Let’s put it this way: even Thariq, a member of the Claude Code team, couldn’t help complaining:
“The post says Claude did it in one shot, but the prompt is 10,000 characters, plus skills, examples, and API keys.”
Even more frightening, blogger Movez then broke down the whole workflow again, and said:
“The prompt is only 10%; the other 90% is the production framework.”
I was stunned, so let me quickly do a flash-step I’ve been practicing at home~
Fortunately, Movez is a good guy, and he also summarized a set of tips for getting Opus 5.5 to do motion design.
What really determines the quality of the final film roughly comes down to these moves.
First, feed it references; don’t just say “make it more high-end.”
You can give it a frame, an entire video, or your own portfolio, and have the model summarize color palette, typography, camera, and transitions into a unified style guide.
That way later visuals won’t each grow in their own direction.
Then, make the animation into a seek(t) time function.
Input the second, and the program can directly calculate the frame for that second. This way the browser can render different segments at the same time, and if one shot fails, you don’t have to rerun the whole video.
For an MV with a song, you also need the model to first analyze BPM, downbeats, and volume peaks, and make a “beat map”:
At second 3, the drum hits; at second 7, the chorus starts; at second 12, the emotion suddenly rises...
Next, cutting shots, popping subtitles, changing colors, and adding sound effects all snap to these time points.
Finally, never render the final version on the first try.
First look at the storyboard and a low-resolution preview, then lay out key frames into one big image to check text, composition, and action.
Pick the three worst problems, fix them, and then render.
One of Movez’s watercolor short films called the model 163 times and ran for nearly 7 hours from start to finish.
So, code-to-direct-output is real.
One-click finished film? Dream on.
Sonnet 5.5’s First Results Leaked
There’s also a latest leak from Sonnet 5.5.
When Opus 5.5 was released, Anthropic casually buried an Easter egg:
“Sonnet 5.5 and Haiku 5.5 will both launch within the next few weeks.”
Now, Sonnet 5.5 is said to be released next week (speculated to be Tuesday local time), and suspected first test results have already surfaced.
The leak account JAZII released a comparison of four models with the same prompt:
GPT-6 Astra, GPT-6 Sol, Claude Opus 5.5, and the allegedly secretly tested Claude Sonnet 5.5, simultaneously generating a pixel animation composed of a moonlit night, cherry blossoms, and a Japanese-style town.
Judging only from the leaked footage, Sonnet 5.5’s frame has denser building layers, richer foreground and background, and is more generous with lights, cherry blossoms, and small elements.
The two OpenAI models’ compositions are relatively flatter, while Opus 5.5 is more restrained and clean.
The leaker’s evaluation showed no mercy:
“This time Sonnet crushed both OpenAI models.”
The result is open to interpretation, but expectations are indeed already there.
If Sonnet really brings Opus’s recent “sudden aesthetic awakening” to a cheaper, more commonly used tier —
Then I’ll just wish it success soon.
Reference links:
[1] https://x.com/anabology/status/2103534482930491441?s=20
[2] https://x.com/_mexicat/status/2103108369569726802?s=20
[3] https://x.com/notjazii/status/2103884167104831573?s=20